Atlas is World Labs' omni world model. It takes text, images, video, and 3D, then generates camera-controlled 1440p video up to a minute, reconstructs scenes from a few photos, and simulates space-time for robotics. Early access.
I’ve been waiting for a world model that actually understands space, not just pretty video. Atlas is that: generate a scene, reconstruct a real room from a couple of photos, then move a camera through it with real geometry.
The demo that got me: two people filmed a scene on a couple of phones, and Atlas turned it into a short film. It generates the parts nobody filmed, freezes time, and lets you reframe from angles that were never shot.
the "reconstructs a scene from a few photos and fills in the parts nobody filmed" part is impressive but also the part that worries me a bit for the robotics use case - if it's confidently generating plausible geometry for spaces it never actually saw, how do you know where the real reconstruction ends and the hallucinated fill-in begins when the output is meant to inform physical navigation?
Love that Atlas treats text, images, video, and 3D as one unified world model instead of separate pipelines, and simulating space-time for robotics is a genuinely forward-looking touch.
Netlify
Hey PH community 👋
Excited to hunt Atlas from World Labs today.
I’ve been waiting for a world model that actually understands space, not just pretty video. Atlas is that: generate a scene, reconstruct a real room from a couple of photos, then move a camera through it with real geometry.
The demo that got me: two people filmed a scene on a couple of phones, and Atlas turned it into a short film. It generates the parts nobody filmed, freezes time, and lets you reframe from angles that were never shot.
https://x.com/davidpantera_/status/2094841083805266401
Eric just used Atlas to generate this 3D reconstruction of his childhood home from only 3 images he grabbed from google street view.
https://x.com/eerac/status/20951...
Early access is open, and this is the model that’ll power Marble next.
Go request access and tell the team what you’d build 🚀
Impressive results. I wonder what is the next step for your team?
Dial
the "reconstructs a scene from a few photos and fills in the parts nobody filmed" part is impressive but also the part that worries me a bit for the robotics use case - if it's confidently generating plausible geometry for spaces it never actually saw, how do you know where the real reconstruction ends and the hallucinated fill-in begins when the output is meant to inform physical navigation?
screenpipe (YC S26)
looks cool - excited to try
Love that Atlas treats text, images, video, and 3D as one unified world model instead of separate pipelines, and simulating space-time for robotics is a genuinely forward-looking touch.